def vgg_a(inputs,
num_classes=1000,
is_training=True,
dropout_keep_prob=0.5,
spatial_squeeze=True,
scope='vgg_a'):
"""Oxford Net VGG 11-Layers version A Example.
Note: All the fully_connected layers have been transformed to conv2d layers.
To use in classification mode, resize input to 224x224.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
spatial_squeeze: whether or not should squeeze the spatial dimensions of the
outputs. Useful to remove unnecessary dimensions for classification.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
with variable_scope.variable_scope(scope, 'vgg_a', [inputs]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with arg_scope(
[layers.conv2d, layers_lib.max_pool2d],
outputs_collections=end_points_collection):
net = layers_lib.repeat(
inputs, 1, layers.conv2d, 64, [3, 3], scope='conv1')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net, 1, layers.conv2d, 128, [3, 3], scope='conv2')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net, 2, layers.conv2d, 256, [3, 3], scope='conv3')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = layers_lib.repeat(net, 2, layers.conv2d, 512, [3, 3], scope='conv4')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
net = layers_lib.repeat(net, 2, layers.conv2d, 512, [3, 3], scope='conv5')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool5')
# Use conv2d instead of fully_connected layers.
net = layers.conv2d(net, 4096, [7, 7], padding='VALID', scope='fc6')
net = layers_lib.dropout(
net, dropout_keep_prob, is_training=is_training, scope='dropout6')
net = layers.conv2d(net, 4096, [1, 1], scope='fc7')
net = layers_lib.dropout(
net, dropout_keep_prob, is_training=is_training, scope='dropout7')
net = layers.conv2d(
net,
num_classes, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='fc8')
# Convert end_points_collection into a end_point dict.
end_points = utils.convert_collection_to_dict(end_points_collection)
if spatial_squeeze:
net = array_ops.squeeze(net, [1, 2], name='fc8/squeezed')
end_points[sc.name + '/fc8'] = net
return net, end_points
def get_slim_arch_bn(inputs, isTrainTensor, num_classes=1000, scope='vgg_16'):
with variable_scope.variable_scope(scope, 'vgg_16', [inputs]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
filters = 64
# Arg scope set default parameters for a list of ops
with arg_scope(
[layers.conv2d, layers_lib.fully_connected, layers_lib.max_pool2d],
outputs_collections=end_points_collection):
net = layers_lib.repeat(
inputs,
2,
layers.conv2d,
filters, [3, 3],
scope='conv1',
weights_regularizer=slim.l2_regularizer(0.01))
bn_0 = tf.contrib.layers.batch_norm(net,
center=True,
scale=True,
is_training=isTrainTensor,
scope='bn1',
decay=0.9)
p_0 = layers_lib.max_pool2d(bn_0, [2, 2], scope='pool1')
net = layers_lib.repeat(
p_0,
2,
layers.conv2d,
filters, [3, 3],
scope='conv2',
weights_regularizer=slim.l2_regularizer(0.01))
bn_1 = tf.contrib.layers.batch_norm(net,
center=True,
scale=True,
is_training=isTrainTensor,
scope='bn2',
decay=0.9)
res_1 = p_0 + bn_1
p_1 = layers_lib.max_pool2d(res_1, [2, 2], scope='pool2')
net = layers_lib.repeat(
p_1,
3,
layers.conv2d,
filters, [4, 4],
scope='conv3',
weights_regularizer=slim.l2_regularizer(0.01))
bn_2 = tf.contrib.layers.batch_norm(net,
center=True,
scale=True,
is_training=isTrainTensor,
scope='bn3',
decay=0.9)
res_2 = p_1 + bn_2
p_2 = layers_lib.max_pool2d(res_2, [2, 2], scope='pool3')
net = layers_lib.repeat(
p_2,
3,
layers.conv2d,
filters, [5, 5],
scope='conv4',
weights_regularizer=slim.l2_regularizer(0.01))
bn_3 = tf.contrib.layers.batch_norm(net,
center=True,
scale=True,
is_training=isTrainTensor,
scope='bn4',
decay=0.9)
res_3 = p_2 + bn_3
p_3 = layers_lib.max_pool2d(res_3, [2, 2], scope='pool4')
last_conv = net = layers_lib.repeat(
p_3,
3,
layers.conv2d,
filters, [5, 5],
scope='conv5',
weights_regularizer=slim.l2_regularizer(0.01))
# Here we have 14x14 filters
net = tf.reduce_mean(net, [1, 2]) # Global average pooling
# add layer with float 32 mask of same shape as global average pooling out
# feed default with ones, leave placeholder
mask = tf.placeholder_with_default(tf.ones_like(net),
shape=net.shape,
name='gap_mask')
net = tf.multiply(net, mask)
net = layers_lib.fully_connected(net,
num_classes,
activation_fn=None,
biases_initializer=None,
scope='softmax_logits')
with tf.variable_scope("raw_CAM"):
w_tensor_name = "vgg_16/softmax_logits/weights:0"
s_w = tf.get_default_graph().get_tensor_by_name(w_tensor_name)
softmax_weights = tf.expand_dims(tf.expand_dims(s_w, 0),
0) # reshape to match 1x1xFxC
# tensor mult from (N x lh x lw x F) , (1 x 1 x F x C)
cam = tf.tensordot(last_conv,
softmax_weights, [[3], [2]],
name='cam_out')
# Convert end_points_collection into a end_point dict.
end_points = utils.convert_collection_to_dict(
end_points_collection)
return net, end_points
def vgg_a(inputs,
num_classes=1000,
is_training=True,
dropout_keep_prob=0.5,
spatial_squeeze=True,
scope='vgg_a'):
"""Oxford Net VGG 11-Layers version A Example.
Note: All the fully_connected layers have been transformed to conv2d layers.
To use in classification mode, resize input to 224x224.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
spatial_squeeze: whether or not should squeeze the spatial dimensions of the
outputs. Useful to remove unnecessary dimensions for classification.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
with variable_scope.variable_scope(scope, 'vgg_a', [inputs]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with arg_scope([layers.conv2d, layers_lib.max_pool2d],
outputs_collections=end_points_collection):
net = layers_lib.repeat(inputs,
1,
layers.conv2d,
64, [3, 3],
scope='conv1')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net,
1,
layers.conv2d,
128, [3, 3],
scope='conv2')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net,
2,
layers.conv2d,
256, [3, 3],
scope='conv3')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = layers_lib.repeat(net,
2,
layers.conv2d,
512, [3, 3],
scope='conv4')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
net = layers_lib.repeat(net,
2,
layers.conv2d,
512, [3, 3],
scope='conv5')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool5')
# Use conv2d instead of fully_connected layers.
net = layers.conv2d(net,
4096, [7, 7],
padding='VALID',
scope='fc6')
net = layers_lib.dropout(net,
dropout_keep_prob,
is_training=is_training,
scope='dropout6')
net = layers.conv2d(net, 4096, [1, 1], scope='fc7')
net = layers_lib.dropout(net,
dropout_keep_prob,
is_training=is_training,
scope='dropout7')
net = layers.conv2d(net,
num_classes, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='fc8')
# Convert end_points_collection into a end_point dict.
end_points = utils.convert_collection_to_dict(
end_points_collection)
if spatial_squeeze:
net = array_ops.squeeze(net, [1, 2], name='fc8/squeezed')
end_points[sc.name + '/fc8'] = net
return net, end_points
def vgg_19(inputs,
num_classes=1000,
is_training=True,
dropout_keep_prob=0.5,
spatial_squeeze=True,
scope='vgg_19'):
"""Oxford Net VGG 19-Layers version E Example.
Note: All the fully_connected layers have been transformed to conv2d layers.
To use in classification mode, resize input to 224x224.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
spatial_squeeze: whether or not should squeeze the spatial dimensions of the
outputs. Useful to remove unnecessary dimensions for classification.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
with variable_scope.variable_scope(scope, 'vgg_19', [inputs]) as sc:
end_points_collection = sc.name + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with arg_scope(
[layers.conv2d, layers_lib.fully_connected, layers_lib.max_pool2d],
outputs_collections=end_points_collection):
feats = []
net = layers_lib.repeat(inputs,
2,
layers.conv2d,
64, [3, 3],
scope='conv1')
feats.append(net)
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net,
2,
layers.conv2d,
128, [3, 3],
scope='conv2')
feats.append(net)
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net,
4,
layers.conv2d,
256, [3, 3],
scope='conv3')
feats.append(net)
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = layers_lib.repeat(net,
4,
layers.conv2d,
512, [3, 3],
scope='conv4')
feats.append(net)
net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
net = layers_lib.repeat(net,
4,
layers.conv2d,
512, [3, 3],
scope='conv5')
feats.append(net)
return feats
with slim.arg_scope(vgg.vgg_arg_scope()):
# logits, end_points= vgg.vgg_16(images, num_classes=3, spatial_squeeze=False)
with tf.variable_scope('vgg_16', 'vgg_16', [images]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
with slim.arg_scope([
layers.conv2d, layers_lib.fully_connected,
layers_lib.max_pool2d
],
outputs_collections=end_points_collection):
dropout_keep_prob = 0.5
is_training = True
net = layers_lib.repeat(images,
2,
layers.conv2d,
64, [3, 3],
scope='conv1')
net = layers.conv2d(net,
number_of_classes, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='fc8')
# Convert end_points_collection into a end_point dict.
end_points = utils.convert_collection_to_dict(
end_points_collection)
logits = net
print(logits.get_shape())
print(labels.shape)
def _create_inference(self,
inputs,
is_training=True,
dropout_keep_prob=0.5):
""" Define the inference model for the network
Args:
Returns:
"""
if self.model_version == 'VGG16':
with tf.variable_scope('vgg_16'):
with framework.arg_scope([
layers.conv2d, layers_lib.fully_connected,
layers_lib.max_pool2d
]):
# From tensorflow.contrib.slim.nets.vgg.vgg_16
net = layers_lib.repeat(inputs,
2,
layers.conv2d,
64, [3, 3],
scope='conv1')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net,
2,
layers.conv2d,
128, [3, 3],
scope='conv2')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net,
3,
layers.conv2d,
256, [3, 3],
scope='conv3')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = layers_lib.repeat(net,
3,
layers.conv2d,
512, [3, 3],
scope='conv4')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
net = layers_lib.repeat(net,
3,
layers.conv2d,
512, [3, 3],
scope='conv5')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool5')
# Custom fc layers to allow variable input image size
net = layers.conv2d(net,
4096,
self.fc6_dims,
padding='VALID',
scope='fc6')
net = layers_lib.dropout(net,
dropout_keep_prob,
is_training=is_training,
scope='dropout6')
net = layers.conv2d(net, 4096, [1, 1], scope='fc7')
net = layers_lib.dropout(net,
dropout_keep_prob,
is_training=is_training,
scope='dropout7')
net = layers.conv2d(net,
self.lbls_dim, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='fc8')
logits = net
elif self.model_version == 'VGG19':
with tf.variable_scope('vgg_19'):
with framework.arg_scope([
layers.conv2d, layers_lib.fully_connected,
layers_lib.max_pool2d
]):
# From tensorflow.contrib.slim.nets.vgg.vgg_19
net = layers_lib.repeat(inputs,
2,
layers.conv2d,
64, [3, 3],
scope='conv1')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net,
2,
layers.conv2d,
128, [3, 3],
scope='conv2')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net,
4,
layers.conv2d,
256, [3, 3],
scope='conv3')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = layers_lib.repeat(net,
4,
layers.conv2d,
512, [3, 3],
scope='conv4')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
net = layers_lib.repeat(net,
4,
layers.conv2d,
512, [3, 3],
scope='conv5')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool5')
# Custom fc layers to allow variable input image size
net = layers.conv2d(net,
4096,
self.fc6_dims,
padding='VALID',
scope='fc6')
net = layers_lib.dropout(net,
dropout_keep_prob,
is_training=is_training,
scope='dropout6')
net = layers.conv2d(net, 4096, [1, 1], scope='fc7')
net = layers_lib.dropout(net,
dropout_keep_prob,
is_training=is_training,
scope='dropout7')
net = layers.conv2d(net,
self.lbls_dim, [1, 1],
activation_fn=None,
normalizer_fn=None,
scope='fc8')
logits = net
return logits
def inference_vanilla_cnn(images, in_training=True):
inputs = images
scope = ''
with variable_scope.variable_scope(scope, 'cifar10', [inputs]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, max_pool2d
with arg_scope([
layers.conv2d, layers.fully_connected, layers_lib.max_pool2d,
layers.batch_norm
],
outputs_collections=end_points_collection):
# Apply specific parameters to all conv2d layers (to use batch norm and relu - relu is by default)
with arg_scope(
[layers.conv2d, layers.fully_connected],
weights_regularizer=regularizers.l2_regularizer(0.0005),
weights_initializer=tf.contrib.layers.xavier_initializer(
uniform=False),
biases_initializer=None,
normalizer_fn=tf.layers.batch_normalization,
normalizer_params=dict(
center=True,
scale=True,
#momentum=BATCHNORM_MOVING_AVERAGE_DECAY, # Decay for the moving averages.
epsilon=0.001, # epsilon to prevent 0s in variance.
training=in_training)):
net = layers_lib.repeat(inputs,
1,
layers.conv2d,
32, [3, 3],
scope='conv1')
_activation_summary(net)
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net,
1,
layers.conv2d,
32, [3, 3],
scope='conv2')
_activation_summary(net)
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net,
1,
layers.conv2d,
32, [3, 3],
scope='conv3')
_activation_summary(net)
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = tf.reshape(net, [net.shape[0], -1])
net = layers.fully_connected(
net,
NUM_CLASSES,
scope='fc4',
activation_fn=None,
normalizer_fn=None,
biases_initializer=tf.constant_initializer(0))
return net
def inference(images, in_training=True):
inputs = images
scope = ''
with variable_scope.variable_scope(scope, 'cifar10', [inputs]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, max_pool2d
with arg_scope([
layers.conv2d, dau_conv2d, layers.fully_connected,
layers_lib.max_pool2d, layers.batch_norm
],
outputs_collections=end_points_collection):
# Apply specific parameters to all conv2d layers (to use batch norm and relu - relu is by default)
with arg_scope(
[layers.conv2d, dau_conv2d, layers.fully_connected],
weights_regularizer=regularizers.l2_regularizer(0.1),
weights_initializer=tf.contrib.layers.xavier_initializer(
uniform=False),
#weights_initializer= lambda shape,dtype=tf.float32, partition_info=None: tf.contrib.layers.xavier_initializer(uniform=False),
biases_initializer=None,
#normalizer_fn=layers.batch_norm,
#normalizer_params={'center': True,
# 'scale': True,
# #'is_training': in_training,
# 'decay': BATCHNORM_MOVING_AVERAGE_DECAY, # Decay for the moving averages.
# 'epsilon': 0.001, # epsilon to prevent 0s in variance.
# 'data_format':'NCHW',
# },
normalizer_fn=tf.layers.batch_normalization,
normalizer_params=dict(
center=True,
scale=True,
#momentum=BATCHNORM_MOVING_AVERAGE_DECAY, # Decay for the moving averages.
epsilon=0.001, # epsilon to prevent 0s in variance.
axis=1,
training=in_training)):
inputs = tf.transpose(inputs, [0, 3, 1, 2])
print("input: ", inputs.shape)
net = layers_lib.repeat(
inputs,
1,
dau_conv2d,
96,
dau_units=(2, 2),
max_kernel_size=25,
mu2_initializer=DAUGridMean(dau_units=(2, 2),
max_value=4,
dau_unit_axis=1),
mu1_initializer=DAUGridMean(dau_units=(2, 2),
max_value=4,
dau_unit_axis=2),
#mu2_initializer=tf.constant_initializer(0),
#mu1_initializer=tf.constant_initializer(0),
dau_unit_border_bound=0.0,
mu_learning_rate_factor=1,
data_format='NCHW',
scope='dau_conv1')
net = layers_lib.max_pool2d(net, [2, 2],
scope='pool1',
data_format="NCHW")
#'''
net = layers_lib.repeat(
net,
1,
dau_conv2d,
96,
dau_units=(2, 2),
max_kernel_size=12,
mu2_initializer=DAUGridMean(dau_units=(2, 2),
max_value=4,
dau_unit_axis=1),
mu1_initializer=DAUGridMean(dau_units=(2, 2),
max_value=4,
dau_unit_axis=2),
#mu2_initializer=tf.constant_initializer(0),
#mu1_initializer=tf.constant_initializer(0),
dau_unit_border_bound=0.0,
mu_learning_rate_factor=1,
data_format='NCHW',
scope='dau_conv2')
net = layers_lib.max_pool2d(net, [2, 2],
scope='pool2',
data_format="NCHW")
net = layers_lib.repeat(
net,
1,
dau_conv2d,
192,
dau_units=(2, 2),
max_kernel_size=12,
mu2_initializer=DAUGridMean(dau_units=(2, 2),
max_value=4,
dau_unit_axis=1),
mu1_initializer=DAUGridMean(dau_units=(2, 2),
max_value=4,
dau_unit_axis=2),
#mu2_initializer=tf.constant_initializer(0),
#mu1_initializer=tf.constant_initializer(0),
dau_unit_border_bound=0.0,
mu_learning_rate_factor=1,
data_format='NCHW',
scope='dau_conv3')
net = layers_lib.max_pool2d(net, [2, 2],
scope='pool3',
data_format="NCHW")
#'''
net = tf.reshape(net, [net.shape[0], -1])
net = layers.fully_connected(
net,
NUM_CLASSES,
scope='fc4',
activation_fn=None,
normalizer_fn=None,
biases_initializer=tf.constant_initializer(0))
return net
def get_slim_arch_bn(inputs, isTrainTensor, num_classes=1000, scope='vgg_16'):
"""
from vgg16 https://github.com/tensorflow/models/blob/master/research/slim/nets/vgg.py
:param inputs:
:param num_classes:
:param scope:
:return:
"""
with variable_scope.variable_scope(scope, 'vgg_16', [inputs]) as sc:
end_points_collection = sc.original_name_scope + '_end_points'
# Collect outputs for conv2d, fully_connected and max_pool2d.
# Arg scope set default parameters for a list of ops
with arg_scope(
[layers.conv2d, layers_lib.fully_connected, layers_lib.max_pool2d],
outputs_collections=end_points_collection):
net = layers_lib.repeat(inputs,
2,
layers.conv2d,
64, [3, 3],
scope='conv1')
net = tf.contrib.layers.batch_norm(net,
center=True,
scale=True,
is_training=isTrainTensor,
scope='bn1')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net,
2,
layers.conv2d,
128, [3, 3],
scope='conv2')
net = tf.contrib.layers.batch_norm(net,
center=True,
scale=True,
is_training=isTrainTensor,
scope='bn2')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net,
3,
layers.conv2d,
256, [3, 3],
scope='conv3')
net = tf.contrib.layers.batch_norm(net,
center=True,
scale=True,
is_training=isTrainTensor,
scope='bn3')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = layers_lib.repeat(net,
3,
layers.conv2d,
512, [3, 3],
scope='conv4')
net = tf.contrib.layers.batch_norm(net,
center=True,
scale=True,
is_training=isTrainTensor,
scope='bn4')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
net = layers_lib.repeat(net,
3,
layers.conv2d,
512, [3, 3],
scope='conv5')
# Here we have 14x14 filters
net = tf.reduce_mean(net, [1, 2]) # Global average pooling
net = layers_lib.fully_connected(net,
num_classes,
activation_fn=None,
biases_initializer=None,
scope='softmax_logits')
# Convert end_points_collection into a end_point dict.
end_points = utils.convert_collection_to_dict(
end_points_collection)
return net, end_points
